HyperplaneGAN: a unified consistent translation framework for facial attribute editing

Facial attribute editing has been obtaining remarkable progress as the rapid development in deep generative models. Existing algorithms can be roughly grouped into two distinct categories: attribute-guided models and exemplar-guided models. These models achieve impressive facial attribute editing results, however, there are some limitations. For example, images generated by current attribute-guided models are lack of diversity and attribute styles are not controllable. For exemplar-guided models, low transfer precision and fidelity of generated images are commonly complained issues. In order to generate high-quality attribute-controllable facial images, we propose a novel unified translation framework called HyperplaneGAN which has following advantages: (1) the proposed model can do both attribute-guided facial editing and exemplar-guided facial editing; (2) by employing latent unit swapping and linear separation constraint for learning pair-wise linearly separable disentangled representations, the model can do flexible and controllable translation; (3) cycle-consistency loss and residual attribute vectors are used to guide the model to manipulate specific attributes precisely while other attributes are kept intact. Substantial experimental results demonstrate that HyperplaneGAN outperforms state-of-the-art models on both attribute-guided facial editing and exemplar-guided facial editing, in terms of quantitative evaluation and qualitative evaluation.